Most citations lead nowhere
Three of the eleven most-cited figures trace to articles with no published method, sample, or definition. They are quoted authoritatively because they have been quoted before.
Research · updated August 2026
Almost every ERP article cites a failure rate between 50 and 75 percent, usually without a source. We went looking for where the number comes from, and the answer is more interesting than the statistic: most of these studies are measuring something buyers would not call failure.
Tell us where your project is. We will tell you honestly which risks apply and which do not.
Where elapsed weeks go in a typical buying process. The two highlighted phases are the ones buyers under-invest in and the ones that predict outcomes.
What we found
This is less a research finding than an act of citation archaeology, and the result changes how the figure should be used.
Three of the eleven most-cited figures trace to articles with no published method, sample, or definition. They are quoted authoritatively because they have been quoted before.
Exceeded budget. Exceeded schedule. Failed to deliver expected benefits. Abandoned. These produce wildly different rates and are quoted interchangeably.
Only two of eleven measured projects that were abandoned. The rest measured overrun against an original estimate, which is a different and much more common thing.
A project that overruns an estimate nobody defended is evidence about the estimate as much as about the project. Several studies do not distinguish the two.
The most useful definition — did it deliver what it was bought for — is also the least measured, because it requires a baseline almost nobody establishes.
Our own sample of stalled projects and every practitioner we know supports the conclusion that these projects go badly at a high rate. The number is soft; the phenomenon is not.
We quote the seventy percent figure on our own site, which is why we went looking for its provenance. If we are going to use a statistic in sales material, we should be able to say where it came from and what it counted.
The honest answer is that the specific number is soft. It comes from a small number of studies with incompatible definitions, several of which are twenty years old, and it has been repeated into apparent authority. We will keep using it, with this page attached, and you should discount it accordingly.
A buyer wants to know one thing: what is the probability that this project delivers the outcome I am buying it for, within a range of cost and time I would have accepted in advance.
Almost no study measures that, because it requires a baseline recorded before the project starts and revisited afterwards — and organisations that have just completed a difficult ERP implementation are not eager to run that assessment. The measurement gap is not accidental.
Most of these studies measure variance against an original estimate. That is worth knowing and it conflates two different failures: a project that went badly, and a project that was estimated badly.
In our own sample of stalled implementations, the original estimates were frequently indefensible on day one — a six-month timeline for a nine-entity migration with unexamined data. Counting that as an implementation failure is true and misses where it actually went wrong.
Rather than a rate, three statements we can support from our own work and would defend. Requirements that were not derived from transactions are the most common cause of stall. Data quality is nearly always worse than the diagnostic phase established. And a project without someone empowered to make cross-departmental decisions in a day will queue until it stalls.
Those are more useful than a percentage because they are actionable. A buyer cannot do anything with a seventy percent failure rate except worry; they can do something about all three of those.
We traced the eleven most frequently cited ERP failure statistics appearing in vendor material, analyst content, and trade press between 2023 and 2026, following citations back to a primary source where one existed.
For each we recorded: the definition of failure used, sample size, sample composition, publication date, and whether a method was published. Three had no traceable primary source. Four dated from before 2010 and described enterprise on-premise implementations that bear limited resemblance to mid-market cloud deployments.
This is citation analysis, not new primary research. It does not establish a better failure rate and it does not claim to. What it establishes is that the widely quoted figures are weaker than their usage implies.
Our own complementary sample — 41 stalled deployments, analysed by primary cause — is published separately and has its own substantial limitations, chiefly that it is self-selecting toward severe failures.
Questions
Requirements from transactions, data examined properly, and a decision-maker with real authority.